CN105631469A - Bird image recognition method by multilayer sparse coding features - Google Patents

Abstract

The invention discloses a bird image recognition method by multilayer sparse coding features. Firstly, a cascaded multilayer sparse coding structure is used on R, G and B channels of a local image block for extracting sparse coding features, wherein each layer of sparse coding structure comprises a feature coding part and a feature maximum value extraction part; and then, in the aspect of output features, the multilayer sparse coding features use a linear kernel for fusion, and an SVM is used as a classifier for classified judgment. In the multilayer sparse coding structure, a local constraint term is added to an optimization objective function, a solved approximate solution for the objective function is iteratively used for sparse coding in the feature coding part, the coding values are then used for minimally reconstructing errors, and a dictionary is updated. According to the bird image recognition method by multilayer sparse coding features, the bird image recognition precision by the system can be greatly improved.

Description

A kind of birds image-recognizing method of multilamellar sparse coding feature

Technical field

The present invention relates to field of image recognition, particularly to the birds image-recognizing method of a kind of multilamellar sparse coding feature.

Background technology

The patent of more existing birds identifications at present, but the patent being identified for birds image does not find. In birds identification, as patent 2009102108999 provides the remote wireless monitoring system based on Acoustic image integrated wild birds identification technology, it is used for long-range ecological monitoring, and described system includes song and records module, video record module, audio, video data processing module, audio/video flow transport module and audiovisual digital file storage module. Patent 2013105810072 provides a kind of portable birds recognition methods based on chirm, the chirm signal gathered is carried out pretreatment by LabVIEW software by this patent, and process the chirm signal after pretreatment by the AOK Time-Frequency Analysis Method of LabVIEW and MATLAB software hybrid programming, finally by MATLAB software processes AOK time-frequency spectrum, it is achieved the extraction of eigenvalue. The eigenvalue of the bird of known bird kind is generated training template through model training and carries out data storage, the eigenvalue of the bird of bird kind to be identified is generated test template through model training, carries out mating the identification realizing birds with training template by test template in conjunction with DTW algorithm. In real life, user's mode such as more often with mobile phone, birds are taken pictures is recorded and identifies. A kind of recognition methods based on birds image it is badly in need of for this.

Summary of the invention

In order to overcome disadvantages mentioned above and the deficiency of prior art, it is an object of the invention to the birds image-recognizing method of a kind of multilamellar sparse coding feature, the accuracy of identification for birds image is high.

The purpose of the present invention is achieved through the following technical solutions:

The birds image-recognizing method of a kind of multilamellar sparse coding feature, comprises the following steps:

S1 trains process

S1.1 collects various birds training images, forms birds training dataset;

S1.2 dictionary learning

S1.2.1 set the object function of dictionary learning as:

$\underset{D,X}{\min }\left|\right|Y-DX|{|}_F^2+\mathrm{\λ}\underset{i=1}{\overset{N}{\Σ}}\underset{j=1,j\≠i}{\overset{N}{\Σ}}|d_i^T{d}_j|+\mathrm{\β}\underset{i=1}{\overset{N}{\Σ}}||e_i\·x_i|{|}^2---\left(1\right)$

Wherein Y is the picture element matrix collection of image block, y_iIt it is the picture element matrix of i-th image block in picture element matrix collection Y; D is the dictionary that need to learn, and dictionary number is N, and dictionary element is d_iAnd d_j; X is the weight coefficient of dictionary, and its element is x_i;It is F norm, e_i=exp ([dist (y_i,d₁),...,dist(y_i,d_N)]^T/ ��), and dist (y_i,d_j) it is y_iWith d_jEuclidean distance, representing matrix dot product, �� is weight, ��, �� be balance two kinds constraint weight coefficients;

S1.2.2 calculates the code coefficient X of input pixel signal matrix stack Y, shown in the fresh target function such as formula (2) obtained, shown in its analytic solutions such as formula (3);

$\underset{D,X}{\min }\left|\right|Y-DX|{|}_F^2+\mathrm{\β}\underset{i=1}{\overset{N}{\Σ}}||e_i\·x_i|{|}^2---\left(2\right)$

$s.t.\∀i,1^T{x}_i=1$

$\left\{\begin{array}{c}{C}_i=(D^T-1y_i^T){(D^T-1y_i^T)}^T\\ x_i=(C_i+{\mathrm{\βdiag}}^2(e\left)\right)\backslash 1\\ x_i=x_i/1^T{x}_i\end{array}\right.---\left(3\right)$

S1.2.3 is after trying to achieve code coefficient X, and the word in dictionary optimizes renewal in order by KSVD algorithm, and object function (1) is updated to:

$\underset{d_m}{\min }\{{\stackrel{\‾}{x}}_m^T{\stackrel{\‾}{x}}_m{d}_m^T{d}_m-2R_m{\stackrel{\‾}{x}}_m+\mathrm{\λ}\underset{j=1,j\≠m}{\overset{N}{\Σ}}|d_j^T{d}_m\left|\right\}---\left(4\right)$

s.t.||d_m||₂=1

WhereinIt is the vector of X m row,It it is the picture element matrix collection Y residual error about m-th word; N is the total number of dictionary;

S1.2.3 constantly iteration realizes step S1.2.1��S1.2.2; Until the dictionary D that each iteration is run no longer changes;

S1.3 utilizes the S1.1 dictionary obtained, and adopts multilamellar sparse coding that birds training image is carried out sparse coding sparse calculation, obtains the output of sparse coding feature;

S1.4 classifier training

The sparse coding feature of step S1.3 gained is exported and sends into linear support vector machine grader, obtain the maximum classification areal model between different classes of birds;

S2 test process

To test image, adopt the method for step S1.3 to obtain the output of sparse coding feature, send into the birds of grader gained of S1.4 different classes of between maximum classification areal model, it is judged that the birds classification output that current test image is corresponding.

Described multilamellar sparse coding is specially two-layer sparse coding, and step is as follows:

In ground floor sparse coding, training image is uniformly divided into two kinds of sizes of 16x16 or 32x32, respectively in the image block of both sizes, utilize the dictionary that S1.1 obtains, carry out sparse coding sparse calculation, obtain the sparse coding coefficient of ground floor, obtain the feature output of ground floor sparse coding;

The feature of ground floor sparse coding is exported the input as second layer sparse coding, in second layer sparse coding, the feature of the sparse coding of ground floor is exported generated two-dimensional matrix, and after it is uniformly divided into 8*8 size, utilize the dictionary that S1.1 obtains, carry out sparse coding sparse calculation, obtain the sparse coding coefficient of the second layer; Obtain the feature output of second layer sparse coding;

The feature of ground floor is exported and the feature output of the second layer carries out cascade, obtain final sparse coding feature output.

The feature output of described acquisition ground floor sparse coding, particularly as follows:

The extraction mode that maximizes utilizing 2*2 size obtains the feature output of ground floor sparse coding.

The feature output of described acquisition second layer sparse coding, particularly as follows:

The extraction mode that maximizes utilizing 2*2 size obtains the feature output of second layer sparse coding.

The described feature output feature of ground floor exported with the second layer carries out cascade, obtains final sparse coding feature output, particularly as follows:

The feature of ground floor is exported and the feature output of the second layer all carries out space delamination sampling, obtain final sparse coding feature output; The processing procedure of space delamination sampling is exactly that image is divided into the subgraph region of 2*2 and 4*4, then again in whole figure, 1/2 subgraph region and 1/4 subgraph region, the rectangular histogram of the feature output of statistics regional, then the feature of regional is exported rectangular histogram and carry out cascade, obtain final sparse coding feature output.

Compared with prior art, the present invention has the following advantages and beneficial effect:

(1) the birds image-recognizing method of the present invention, first at the R of topography's block, utilizing cascade multilamellar sparse coding structure extraction sparse coding feature in G, B triple channel, every layer of sparse coding structure all comprises feature coding part and profile maxima extracts two parts; Then on output characteristic, multilamellar sparse coding feature uses linear kernel to merge, and uses SVM to carry out classification judgement as grader. The present invention adopts multilamellar sparse coding structure, obtained by spatial decimation and multiple dimensioned merge with spatial distribution rank, to extract the diagnostic characteristics of enough discriminations as much as possible, maximize the class inherited of epicritic vision target, improve the system accuracy of identification for birds image.

(2) present invention is in multilamellar sparse coding structure, optimization object function adds a local restriction item, the approximate solution using tried to achieve object function in feature coding part iteratively carries out sparse coding, then these encoded radios are used to minimize reconstructed error, update dictionary, emphasize local smoothing method character in an encoding process.

(3) the birds image-recognizing method of the present invention can rely on merely the information of image that wherein birds classification is judged, method is convenient easily to be realized.

Accompanying drawing explanation

Fig. 1 is the schematic flow sheet of the birds image-recognizing method of the multilamellar sparse coding feature of embodiments of the invention.

Fig. 2 is the multilamellar sparse coding process schematic of embodiments of the invention.

Detailed description of the invention

Below in conjunction with embodiment, the present invention is described in further detail, but embodiments of the present invention are not limited to this.

Embodiment

As it is shown in figure 1, the birds image-recognizing method of the multilamellar sparse coding feature of the present embodiment, including following;

S1 trains process

S1.1 collects various birds training images, forms birds training dataset;

The present invention selects California Institute of Technology 200 class birds data set, and this data set comprises the view data of 200 class difference birds, and each classification picture number has more than 300.

S1.2 dictionary learning

S1.2.1

Before carrying out classifier training, it is necessary to set up sparse coding dictionary. In traditional sparse coding dictionary learning process, owing to some image block can substantial amounts of repeat, therefore, when allusion quotation is practised handwriting in randomly drawing sample training, these significantly high image blocks of repeatability will be selected with significantly high probability, thus cause final study to dictionary Expired Drugs occurs. Simultaneously in the process of study dictionary, some have the visually bigger image block of texture difference can use same atoms in dictionary, and therefore dictionary learning process is not sufficiently stable. In order to make dictionary learning process have the character of local smoothing method, simultaneously also can fast coding, the present invention at object function plus a local smoothing method bound term, and iteration go optimize new object function.

The present embodiment set the object function of dictionary learning as:

$\underset{D,X}{\min }\left|\right|Y-DX|{|}_F^2+\mathrm{\λ}\underset{i=1}{\overset{N}{\Σ}}\underset{j=1,j\≠i}{\overset{N}{\Σ}}|d_i^T{d}_j|+\mathrm{\β}\underset{i=1}{\overset{N}{\Σ}}||e_i\·x_i|{|}^2---\left(1\right)$

Wherein Y is that (the present embodiment is that training data is concentrated image have employed the image array of 20000 parts of 16*16 sizes, y at random for the picture element matrix collection of image block_iIt is the picture element matrix of wherein i-th image block); D is the dictionary that need to learn, and dictionary number is N, and dictionary element is d_iAnd d_j; X is the weight coefficient of corresponding dictionary, and its element is x_i;It is F norm, e_i=exp ([dist (y_i,d₁),...,dist(y_i,d_N)]^T/ ��), and dist (y_i,d_j) it is y_iWith d_jEuclidean distance, representing matrix dot product, �� is weight (the present embodiment sets it as 0.1), ��, �� be balance two kinds constraint weight coefficients (it is 0.003 that the present embodiment sets �� as 0.01, ��). Owing to relating to solving of two variablees of X and D, the method for iteration optimization is adopted to implement for this.

S1.2.2 calculates the code coefficient X of input pixel signal matrix stack Y, and therefore the Section 2 in formula (1) can be omitted, shown in the fresh target function such as formula (2) obtained, shown in its analytic solutions such as formula (3);

$\underset{D,X}{\min }\left|\right|Y-DX|{|}_F^2+\mathrm{\β}\underset{i=1}{\overset{N}{\Σ}}||e_i\·x_i|{|}^2---\left(2\right)$

$s.t.\∀i,1^T{x}_i=1$

$\{\begin{array}{c}{C}_i=(D^T-1y_i^T){(D^T-1y_i^T)}^T\\ x_i=(C_i+{\mathrm{\βdiag}}^2(e))\backslash 1\\ x_i=x_i/1^T{x}_i\end{array}---\left(3\right)$

S1.2.3 is after trying to achieve code coefficient X, and the word in dictionary optimizes renewal in order by KSVD algorithm, and by the constant part in removable (1), object function (1) is updated to:

$\underset{d_m}{\min }\{{\stackrel{\‾}{x}}_m^T{\stackrel{\‾}{x}}_m{d}_m^T{d}_m-2R_m{\stackrel{\‾}{x}}_m+\mathrm{\λ}\underset{j=1,j\≠m}{\overset{N}{\Σ}}|d_j^T{d}_m|+\mathrm{\β}\underset{i=1}{\overset{N}{\Σ}}\exp \left(2dist\right(y_i,d_m)/\mathrm{\σ})*{\stackrel{\‾}{x}}_{mi}^2*{\mathrm{\δ}}_{mi}\}---\left(4\right)$

s.t.||d_m||₂=1

WhereinIt is the vector of the m row of X,Being the matrix residual error about m-th word, the total number of dictionary is N,Be in X (m, i) element, ifThen ��_mi=1, otherwise ��_mi=0. In order to solve formula (4), it is possible to find 0 < exp (2dist (y of wherein the 3rd bound term_i,d_m)/��) < 1,Both products make Section 3 be almost 0, it is possible to this omitted; From another perspective, this actual meaning is introduced into the local restriction in cataloged procedure, and with new middle maintenance, this is only the d making to solve out at dictionary_mCloser to training sample, this it is not necessarily the case.

S1.2.3 constantly iteration realizes step S1.2.1��S1.2.2; Until the dictionary D that each iteration is run no longer changes;

S1.3 utilizes the S1.1 dictionary obtained, and adopts multilamellar sparse coding that birds training image is carried out sparse coding sparse calculation, obtains the output of sparse coding feature, as shown in Figure 2:

In ground floor sparse coding, training image is uniformly divided into two kinds of sizes of 16x16 or 32x32, respectively in the block of both image sizes, utilize the dictionary that S1.1 obtains, carry out sparse coding sparse calculation, obtain the sparse coding coefficient of ground floor, utilize the extraction mode that maximizes of 2*2 size to obtain the feature output of ground floor sparse coding;

The feature of ground floor sparse coding is exported the input as second layer sparse coding, in second layer sparse coding, the feature of the sparse coding of ground floor is exported generated two-dimensional matrix, and after it is uniformly divided into 8*8 size, utilize the dictionary that S1.1 obtains, carry out sparse coding sparse calculation, obtain the sparse coding coefficient of the second layer; The extraction mode that maximizes utilizing 2*2 size obtains the feature output of second layer sparse coding;

The feature of ground floor is exported and the feature output of the second layer all carries out space delamination sampling, obtain final sparse coding feature output; The processing procedure of space delamination sampling is exactly that image is divided into the subgraph region of 2*2 and 4*4, then in whole figure, 1/2 subgraph region and 1/4 subgraph region, the rectangular histogram of the feature output of statistics regional, then the feature of regional is exported rectangular histogram and carry out cascade, obtain final sparse coding feature output.

S1.4 classifier training

The sparse coding feature of step S1.3 gained exports the linear support vector machine grader of feeding, and (support vector machine grader principle is referred to document Chang, Chih-Chung; Lin, Chih-Jen, 2011, LIBSVM:Alibraryforsupportvectormachines, ACMTransactionsonIntelligentSystemsandTechnology), obtain the maximum classification areal model between different classes of birds;

S2 test process

To test image, adopt the method for step S1.3 to obtain the output of sparse coding feature, send into the birds of grader gained of S1.4 different classes of between maximum classification areal model, it is judged that the birds classification output that current test image is corresponding.

Above-described embodiment is the present invention preferably embodiment; but embodiments of the present invention are also not restricted by the embodiments; the change made under other any spirit without departing from the present invention and principle, modification, replacement, combination, simplification; all should be the substitute mode of equivalence, be included within protection scope of the present invention.

Claims (5)

1. the birds image-recognizing method of a multilamellar sparse coding feature, it is characterised in that comprise the following steps:

S1 trains process

S1.1 collects various birds training images, forms birds training dataset;

S1.2 dictionary learning

S1.2.1 set the object function of dictionary learning as:

$\underset{D,X}{min}\left|\right|Y-DX|{|}_F^2+\mathrm{\&lambda;}\underset{i=1}{\overset{N}{\&Sigma;}}\underset{j=1,j\&NotEqual;i}{\overset{N}{\&Sigma;}}|d_i^T{d}_j|+\mathrm{\&beta;}\underset{i=1}{\overset{N}{\&Sigma;}}||e_i\&CenterDot;x_i|{|}^2---\left(1\right)$

Wherein Y is the picture element matrix collection of image block, y_iIt it is the picture element matrix of i-th image block in picture element matrix collection Y; D is the dictionary that need to learn, and dictionary number is N, and dictionary element is d_iAnd d_j; X is the weight coefficient of dictionary, and its element is x_i;It is F norm, $e_i=\exp ({\&lsqb;dist(y_i,d_1),...,dist(y_i,d_N)\&rsqb;}^T/\mathrm{\&sigma;}),$ And dist (y_i,d_j) it is y_iWith d_jEuclidean distance, representing matrix dot product, �� is weight, ��, �� be balance two kinds constraint weight coefficients;

S1.2.2 calculates the code coefficient X of input pixel signal matrix stack Y, shown in the fresh target function such as formula (2) obtained, shown in its analytic solutions such as formula (3);

$\underset{D,X}{min}\left|\right|Y-DX|{|}_F^2+\mathrm{\&beta;}\underset{i=1}{\overset{N}{\&Sigma;}}||e_i\&CenterDot;x_i|{|}^2---\left(2\right)$

$s.t.\&ForAll;i,1^T{x}_i=1$

$\left\{\begin{array}{c}{C}_i=(D^T-1y_i^T){(D^T-1y_i^T)}^T\\ x_i=(C_i+{\mathrm{\&beta;diag}}^2(e))\backslash 1\\ x_i=x_i/1^T{x}_i\end{array}\right.---\left(3\right)$

S1.2.3 is after trying to achieve code coefficient X, and the word in dictionary optimizes renewal in order by KSVD algorithm, and object function (1) is updated to:

$\underset{d_m}{min}\{{\stackrel{\&OverBar;}{x}}_m^T{\stackrel{\&OverBar;}{x}}_m{d}_m^T{d}_m-2R_m{\stackrel{\&OverBar;}{x}}_m+\mathrm{\&lambda;}\underset{j=1,j\&NotEqual;m}{\overset{N}{\mathrm{\&Sigma;}}}|d_j^T{d}_m\left|\right\}---\left(4\right)$

s.t.||d_m||₂=1

WhereinIt is the vector of X m row,It it is the picture element matrix collection Y residual error about m-th word; N is the total number of dictionary;

S1.2.3 constantly iteration realizes step S1.2.1��S1.2.2; Until the dictionary D that each iteration is run no longer changes;

S1.3 utilizes the S1.1 dictionary obtained, and adopts multilamellar sparse coding that birds training image is carried out sparse coding sparse calculation, obtains the output of sparse coding feature;

S1.4 classifier training

The sparse coding feature of step S1.3 gained is exported and sends into linear support vector machine grader, obtain the maximum classification areal model between different classes of birds;

S2 test process

To test image, adopt the method for step S1.3 to obtain the output of sparse coding feature, send into the birds of grader gained of S1.4 different classes of between maximum classification areal model, it is judged that the birds classification output that current test image is corresponding.

2. the birds image-recognizing method of multilamellar sparse coding feature according to claim 1, it is characterised in that described multilamellar sparse coding is specially two-layer sparse coding, and step is as follows:

In ground floor sparse coding, training image is uniformly divided into two kinds of sizes of 16x16 or 32x32, respectively in the image block of both sizes, utilize the dictionary that S1.1 obtains, carry out sparse coding sparse calculation, obtain the sparse coding coefficient of ground floor, obtain the feature output of ground floor sparse coding;

The feature of ground floor sparse coding is exported the input as second layer sparse coding, in second layer sparse coding, the feature of the sparse coding of ground floor is exported generated two-dimensional matrix, and after it is uniformly divided into 8*8 size, utilize the dictionary that S1.1 obtains, carry out sparse coding sparse calculation, obtain the sparse coding coefficient of the second layer; Obtain the feature output of second layer sparse coding;

The feature of ground floor is exported and the feature output of the second layer carries out cascade, obtain final sparse coding feature output.

3. the birds image-recognizing method of multilamellar sparse coding feature according to claim 2, it is characterised in that the feature output of described acquisition ground floor sparse coding, particularly as follows:

The extraction mode that maximizes utilizing 2*2 size obtains the feature output of ground floor sparse coding.

4. the birds image-recognizing method of multilamellar sparse coding feature according to claim 3, it is characterised in that the feature output of described acquisition second layer sparse coding, particularly as follows:

The extraction mode that maximizes utilizing 2*2 size obtains the feature output of second layer sparse coding.

5. the birds image-recognizing method of multilamellar sparse coding feature according to claim 2, it is characterised in that the described feature output feature of ground floor exported with the second layer carries out cascade, obtains final sparse coding feature output, particularly as follows:

The feature of ground floor is exported and the feature output of the second layer all carries out space delamination sampling, obtain final sparse coding feature output; The processing procedure of space delamination sampling particularly as follows: be divided into the subgraph region of 2*2 and 4*4 by image, then in whole figure, 1/2 subgraph region and 1/4 subgraph region, the rectangular histogram of the feature output of statistics regional, then the feature of regional is exported rectangular histogram and carry out cascade, obtain final sparse coding feature output.